Systematic Calibration of A Convection-Resolving Model: Application over Tropical Atlantic

Non-hydrostatic km-scale weather and climate models are promising in simulating clouds, especially convective ones. However, even km-scale models need to parameterize some physical processes and are thus subject to the corresponding uncertainty of parameters. Systematic calibration has the advantage of improving model performance with transparency and reproducibility, thus benefiting model intercomparison projects, process studies, and climate-change scenario simulations. 

In this paper, the regional atmospheric climate model COSMO v6 is systematically calibrated over the Tropical South Atlantic. First, the parameters' sensitivities are evaluated with respect to a set of validation fields (outgoing longwave radiation (OLR), outgoing shortwave radiation (OSR) and latent heat flux (LHFL)). Five of the most sensitive parameters are chosen for calibration. The objective calibration then closely follows the methodology of Bellprat et al. (2016). This includes simulations considering the interaction of all pairs of parameters and the exploitation of a quadratic-form metamodel to emulate the simulations. In the current set-up with 5 parameters, 50 simulations are required to build the metamodel. Then Latin hypercube sampling is applied and the set of parameters with the best performance score is chosen as the optimal parameter set. The model is calibrated for the year 2016 and validated in 2013. And  the optimal parameter setting lead to significant improvements for both years, especially for OSR, which is closely related to low clouds. More specifically, the domain annual mean OSR bias is reduced from 40 to 13.5 Wm^-2. Moreover, when we apply the optimal setting over a larger domain with a slightly higher resolution (from 4km to 3km) in 2006, the optimal setting still works, especially for OSR and for the calibrated domain. 

The results thus show that parameter calibration is a useful and efficient tool for model improvement. We will also discuss potential limitations and highlight how the approach could be extended to global atmospheric models. Calibrating over a larger domain might help improve the overall performance, but would potentially also lead to compromises among different regions and variables, and require more computational resources.